Generative data assimilation
Combining score-based generative models and Bayesian filtering to infer high-dimensional states from sparse, noisy observations.
Latent-EnSF · LD-EnSF · LD-EnFFComputer Science Ph.D. candidate · Georgia Tech
Phillip Si
Learning from uncertainty.
Modeling the physical world.
I work on probabilistic machine learning, generative data assimilation, and differentiable surrogate models. My research brings sparse observations and dynamical models together to build efficient, uncertainty-aware digital twins.
At Georgia Tech, I’m advised by Peng Chen and Chao Zhang. I also conduct research at Oak Ridge National Laboratory, following work at Argonne National Laboratory.
What I work on
Making sense of complex systems.
From a few measurements to a fuller picture of how the world evolves.
Differentiable dynamics
Learning compact surrogate models for efficient forecasting, joint state–parameter estimation, and model predictive control.
LEVDA · Neural ODEs · Flow controlAI digital twins
Bringing simulation, observations, and learning together for flood prediction, weather, and unsteady fluid dynamics.
CLDNet · C-STRIDE · Flood forecastingSelected work
Recent publications
LEVDA: Latent Ensemble Variational Data Assimilation via Differentiable Dynamics
A differentiable latent ensemble-variational smoother for joint state–parameter estimation from sparse, irregular observations.
Toward AI-driven digital twins for metropolitan floods: A conditional latent dynamics network surrogate of the shallow water equations
Rainfall-driven latent neural dynamics and terrain-conditioned decoding for efficient, high-resolution basin-wide flood forecasts.
LD-EnSF: Latent Dynamics with Ensemble Score Filters for Fast Data Assimilation with Sparse Observations
Latent dynamics and history-aware observation encoding accelerate forecasting and assimilation for complex dynamical systems.
Latent-EnSF: A Latent Ensemble Score Filter for High-Dimensional Data Assimilation with Sparse Observation Data
Coupled variational autoencoders and ensemble score filtering for high-dimensional systems with highly sparse observations.
Along the way
Research experience
May 2026 – Present
Oak Ridge National Laboratory
Graduate Researcher
Sparse-sensor estimation, latent forecasting, and control for unsteady fluid flow.
June 2025 – May 2026
Argonne National Laboratory
Research Aide
Rainfall-driven neural surrogates and digital twins for basin-wide flood prediction.
August 2023 – Present
Georgia Institute of Technology
Graduate Researcher
Generative data assimilation and differentiable latent models for physical systems.
Previously: machine learning graduate coursework at Carnegie Mellon, and a B.A. in Computer Science & Mathematics at Cornell, where I worked with Volodymyr Kuleshov.
Beyond research
Another kind
of experiment.
Away from research, I enjoy cooking and gaming. Here are a few things I’ve made in the kitchen.
Visit the kitchen
Pistachio & herb-crusted lamb